Fetal heart rate estimation in phonocardiograms using deep learning and digital signal processing

Abstract

Fetal heart rate monitoring through phonocardiograms enables non-invasive assessment of fetal health, supporting the early detection of potential complications during pregnancy. However, accurate frequency estimation is hindered by various noise sources, such as maternal heart sounds and ambient interference. This study proposes a lightweight 1D-CNN regression model for fetal heart rate estimation, leveraging MFCC and Delta-MFCC coefficients extracted from audio segments. A simple but effective data augmentation technique was also employed to mitigate the scarcity of labeled data. Experiments conducted on the SUFHSDB dataset yielded a mean absolute error of 3.94±0.41 bpm. The results suggest that deep learning-based approaches, especially when combined with data augmentation, are promising alternatives for fetal heart rate estimation, potentially reducing reliance on traditional signal-processing pipelines.